Propose an action from an agent, bind it to a task contract, and watch it run through the closed loop: policy evaluation, approval gates, bounded execution, and post-condition verification. Everything is captured in an immutable audit ledger.
{
"op": "train.run",
"channel": "gpu_training",
"risk_class": "CONSEQUENTIAL_WRITE",
"targets": [
"cluster://gpu-node-4"
],
"expected_effect": "Emit candidate checkpoint; retain if BPB improves ≥0.001",
"policy_hints": [
"depth ≤ 12",
"memory ≤ 40GB",
"time ≤ 300s"
]
}